Use it in the instruction namer to make instruction names more stable
across unrelated changes to the toolchain or the prelude.
---------
Co-authored-by: Dana Jansens <danakj@orodu.net>
Previously Collect() was used for types that implemented
CollectMemUsage() but otherwise Add() was used. This required the caller
to think about the type of the field and know/decide which method to
use.
Now, the caller always uses Collect() unless they are adding specific
byte values, in which case Add is used. Typically then, Add will only be
used to implement the CollectMemUsage() function.
To do this we require all Collect() methods to be templates so that they
all be a single overload set. The Collect on BumpPtrAllocator is
converted to a template that checks
`std::same_as<llvm::BumpPtrAllocator, T>`.
This is in anticipation of making the integer value store be customized
heavily. I'd like to extract it from the common code when doing that, so
first disentangling them here without any intended change in
functionality or behavior to enable that.
I've tried to update `#include`s to be as minimal as I can and added a
few missing includes spotted in the process.
I've split the test for value store to include what was easy focused on
just the value store templates rather than the unified shared value
stores.
This might surface some opportunities for adding more tests, but for
this PR, just doing the minimal restructuring.
When profiling, these jumped out as good inline candidates that happened
to be out-of-line, this just moves them inline so that they're
available. I think as much as 10% improvement in check-phase from this,
but I haven't run detailed before/after measurements as these changes
seemed minimally disruptive.
Also switched from `CHECK` to `DCHECK` in one place that seems
especially hot and where the check itself seems reasonable to only do in
debug builds. Left a comment since we rarely need to remove these any
more.
This switches `DCHECK` and `FATAL` as well.
The goal is to reduce the code size impact of these assertions so that
we can keep more of them enabled. Currently, the largest cost I see from
`CHECK` is not the actual check or the cold code itself, but actually
the failure to inline trivial functions due to the presence of the cold
code. This means that our goal isn't to reduce apparent code size in the
final binary but the LLVM IR cost assessed for these routines in the
inliner, which closely correlates with code size but is a bit different.
As discussed in #4283, experimentation shows that a single function call
with a minimal number of arguments is the lowest cost model for these.
This is easily achieved with a format-string API that internally uses
`llvm::formatv`. This PR is essentially the `CHECK` version of #4283.
However, the check macros are substantially harder to make work with
both format strings and streaming because they also take a condition.
Also, unexpectedly, I was very successful at devising a regular
expression based automated rewrite from the streaming to the format
string form with only low 10s of manual fixes. This includes compacting
strings broken up across lines, etc. Given how well that went, I've
prepared this PR which just directly switches to the format string API
and migrate everything to use it.
One nice side-effect is that the format string approach ends up greatly
simplifying the implementation here as well.
This is ... *shockingly* effective. Parsing speeds up by more than 3%
with just this change. And checking speeds up by **8%** with this change
alone:
```
BM_CompileAPIFileDenseDecls<Phase::Parse>/256 86.3µs ± 1% 82.9µs ± 1% -3.94% (p=0.000 n=17+19)
BM_CompileAPIFileDenseDecls<Phase::Parse>/1024 431µs ± 1% 415µs ± 1% -3.76% (p=0.000 n=18+19)
BM_CompileAPIFileDenseDecls<Phase::Parse>/4096 1.77ms ± 1% 1.71ms ± 1% -3.18% (p=0.000 n=18+19)
BM_CompileAPIFileDenseDecls<Phase::Parse>/16384 7.44ms ± 1% 7.17ms ± 2% -3.56% (p=0.000 n=18+20)
BM_CompileAPIFileDenseDecls<Phase::Parse>/65536 30.7ms ± 1% 29.7ms ± 1% -3.15% (p=0.000 n=18+20)
BM_CompileAPIFileDenseDecls<Phase::Parse>/262144 131ms ± 1% 127ms ± 1% -2.81% (p=0.000 n=18+18)
BM_CompileAPIFileDenseDecls<Phase::Check>/256 878µs ± 2% 800µs ± 1% -8.91% (p=0.000 n=19+20)
BM_CompileAPIFileDenseDecls<Phase::Check>/1024 1.88ms ± 2% 1.72ms ± 1% -8.56% (p=0.000 n=19+20)
BM_CompileAPIFileDenseDecls<Phase::Check>/4096 5.78ms ± 2% 5.28ms ± 1% -8.70% (p=0.000 n=20+18)
BM_CompileAPIFileDenseDecls<Phase::Check>/16384 21.9ms ± 1% 20.1ms ± 1% -8.02% (p=0.000 n=18+20)
BM_CompileAPIFileDenseDecls<Phase::Check>/65536 90.4ms ± 2% 83.1ms ± 1% -8.04% (p=0.000 n=19+20)
BM_CompileAPIFileDenseDecls<Phase::Check>/262144 381ms ± 2% 352ms ± 1% -7.79% (p=0.000 n=19+19)
```
---------
Co-authored-by: Richard Smith <richard@metafoo.co.uk>
Co-authored-by: josh11b <15258583+josh11b@users.noreply.github.com>
Most of these are about enabling inlining, in a couple of cases moving
code to a header and throughout switching to `CARBON_DCHECK`. The code
size of `CARBON_CHECK` seems to make inliing quite unreliable. I'm going
to think about whether there are ways to improve this, but a reasonably
small number of these seem worth switching for now to get some compile
time savings.
Also moves VLOG out of the hot path which helps a bit as well.
All combined, this net a bit over 10%, although it varies a bit exactly
how much. We're now pretty consistently over 800k lines/second for check
in the compilation benchmark for files >=4k lines, which makes me happy.
That's remarkably close to our original target.
Not really planning to keep optimizing here, just was glancing at the
profile and many of these stood out to me and were easy to fix.
This uses a heuristic reserve to greatly reduce hashtable growth of the
identifiers hashtable. The design of the hashtable itself is optimized
around compact memory use and is especially slow to grow and so this has
an outsized impact.
The heuristic was computed using `scripts/source_stats.py` and looking
at C++ codebases. We may want to periodically re-evaluate it as Carbon
code emerges and we have better data on its distributions of tokens.
This also required fixing the `Reserve` method on `CanonicalValueStore`
that wasn't actually used anywhere and so didn't even compile correctly.
I added it to the relevant unit test so it is at least compiled locally
to its definition.
This undoes a previous change to unify them, and I think at my advice.
=[ Sorry about that, I think I was just wrong.
Specifically, I think I had suggested that it would be more efficient to
have a single shared hashtable of strings. The more I look at profiles
of the toolchain, the less likely that seems. Specifically for
identifiers and string literals it seems especially problematic.
Using a single, joint hashtable is likely a good idea when all of the
different querying code paths are equally likely, the strings follow the
same distribution of sizes, and either there is no clustering of access
to different sets of strings or none of the sets are meaningfully small
enough to fit into a lower level of resident cache.
I think essentially none of these predicates actually hold for
identifiers vs. string literals:
- Identifiers are *much* more hot
- They have wildly different size distributions.
- The access patterns are very clustered
Sorry for the misleading advice on that one.
While splitting them, I've worked to simplify the code a bit by building
a way to have the `StringRef` holding canonical value stores not require
specializations, and so we get a pretty large code cleanup in the
process here.
This works to leverage the capabilities of the hashtable as much as
possible, for example using the key context in the value stores.
However, there may still be opportunities to refactor more deeply and
use the functionality even better. Hopefully this is at least
a reasonable start and gets us a clean baseline.
On an Arm M1, this is a 15% improvement on my large lexing stress test,
but ends up a wash on my x86-64 server. This is a smaller benefit than
I expected, and it's because we're using a set-of-IDs and looking up
values with a key context for things like identifiers. This pattern has
a surprising tradeoff. The new hashtable uses significantly less memory,
a 10% peak RSS reduction just from the hashtable change. But indirecting
through the vector of values makes growing the hashtable dramatically
less cache-friendly: it causes growth to randomly access every key when
rehashing. On x86, everything gained by the faster hashtable is lost in
even slower growth. And even on Arm, this eats into the benefits.
But I have a plan to tweak how identifiers specifically work to avoid
most of the growth, and so I suspect this is the right tradeoff on the
whole. It gives us significant working set size reduction and we can
likely avoid the regressed operation (growth with rehash) in most cases
by clever reserving and if necessary by adding a hash caching layer to
the table infrastructure.
---------
Co-authored-by: Jon Ross-Perkins <jperkins@google.com>
Switch from recursing into non-canonical instruction fields to
separately canonicalizing those fields. This means we now form canonical
`InstBlockId`s, `TypeBlockId`s, `IntId`s, `FloatId`s, and `BindNameId`s
at least in the cases when they're referenced by a constant instruction.
This reduces the overall runtime for @chandlerc's 10MLoC example by
27.5% on my machine.
Adds support for unary `-` and binary `+`, `-`, `*`, `/` for floating
point types.
Real literals are now transformed to `llvm::APFloat`s during the check
phase into the `FloatLiteral` instruction.
This PR likely collides a bit with #3892 and might need to be updated
when that one is merged.
Note, I'm annotating the lookup partly so that the reason the conflict
comes up is clear, partly so that there's actually a diagnostic line
associated with the root cause as more tests get packed into a single
file.
This is primarily being done for performance reasons, removing hash
lookups. The increased memory consumption is accepted.
Refactors LexicalLookup out to its own structure.
Namespaces are copied, which means also adding their name to the
underlying instruction. It happened not to be done previously; the name
was only in name lookup.
Since the only import supported right now is the default import,
functionality is limited; in the future I'll need to deal with namespace
vs package conflicts.
Tests of namespace imports are under "namespace" -- I figured this would
be best for scaling as more instructions get support.
This also improves some debugging-related output that I was trying to
use while trying to build the support.
---------
Co-authored-by: Richard Smith <richard@metafoo.co.uk>
We have `StringLiteral`s in multiple other `Carbon` sub-namespaces.
Rename to a more specific name to avoid collisions.
We should likely also rename `Carbon::IntId` -> `Carbon::IntValueId` and
`Carbon::RealId` -> `Carbon::RealValueId`, but this collision is
prioritized because it was blocking work on typed parse nodes which
introduces a `Carbon::Parse::StringLiteralId`.
Adds a `LazyImportRef` instruction. Versus `CrossRef`, this is intended
to represent an instruction which cannot be used directly, and must be
replaced when it comes up due to name lookup. The intent is to use this
to avoid recursive loading of imported IR instructions.
Note, under this model, when `ResolveIfLazyImportRef` is called, it
essentially needs to load both inst and type information to a sufficient
point where any further attempts would hit name lookup again. That will
probably be complex, and the current implementation is just touching the
surface of the issue. I was heading down this route because it would
mean we have a limited number of points that need to consider whether
they're going to talk about a `LazyImportRef`.
I'm considering whether `CrossRef` should be dropped in favor of more
specific `Builtin` special-casing, due to the divergence of desired
behaviors. This could mean dropping the `builtins` IR since it's not
looking useful right now.
Modify `NameScope` to track whether the scope is associated with a load
error. This is to handle cases where one or more imports failed, so we
do not want to issue warnings for related scopes.
The 0-size on `ValueStore` comes up due to the changes to `NameScope`,
which make it too large for the default handling. After discussion with
zygoloid, the thought was we might want to try reserving a roughly
correct value based on parse node counts, but the stack default wasn't
buying much.
Fixes a bug where the implicit import used the package name instead of
the invalid identifier.
There's a trade-off here of explicitness in the use versus repetition,
but I'm hoping the Id offers sufficient info (also, some Ids already
relied on this, so this builds consistency). The forward declarations
I'm mixed on, but they are difficult to avoid if heading down this route
due to interdependencies between ids and types which contain ids.
This reflects how we're naming classes that derive from these classes,
and matches usage for each existing `Id` and `Index` type, except:
- `Parse::NodeId` previously inherited from `ComparableIndexBase`, and
is no longer comparable.
- `SemIR::MemberIndex` previously inherited from `IndexBase`, and is now
comparable.
Making `Parse::NodeId` non-comparable reflects that it's intended to be
an opaque identifier for a node and that the ordering is an
implementation detail rather than part of the intended public interface.
`PostorderIterator` and `SiblingIterator` still rely on the numerical
meaning of `NodeId`s, but that's OK since they're part of the node
implementation.
I was suggesting this because `FloatingPoint` is pretty long. `int` and
`float` should be familiar abbreviations. `unsigned` should be familiar
to developers too, but `UnsignedInt` still feels usefully clearer for
the additional chars.
These are manual fixes; mostly from clang-tidy, some from clangd (which
notes unused includes).
In typed_insts, adding inlline due to misc-definitions-in-headers. Per
discussion, clang-tidy is wrong, but inline silences it.
For parameter name skew in definition versus declaration, I'm just using
the name from the definition.
---------
Co-authored-by: Richard Smith <richard@metafoo.co.uk>
Add a `NameId` that is effectively just a wrapper around a `StringId`,
with
some additional predefined values for names that don't correspond to
strings, such as the name of `self` or the function's return slot.
Following up on discussion yesterday regarding this split.
Note, I'm expecting #3341 to do IdentifierId -> NameId in SemIR. It
might be worth adding NameId creation directly to StringStore if you're
content with this setup though.
Advantages:
- Allows lexing/parsing in parallel, since they are modifying fully
separate ValueStores.
- Allows SemIR to reliably be stored hermetically.
Disadvantages:
- Creates overlapping storage of duplicate strings when multiple files
are compiled together.
- Prevents Ids from being uniquely compared cross-file.
Per discussion, the decision is that the advantages are more important.
The looser ownership remains because both SemIR checking and
metaprogramming may still generate things we would want to deduplicate.
It would be somewhat odd if TokenizedBuffer owned something that
checking modified.
Provides an adapter for the llvm::yaml API because it otherwise needs a
bunch of const/non-const definitions, and the traits are difficult to
diagnose issues with. The current approach is pretty simple to use, even
if it's not super efficient (which, yaml output is more of a debugging
thing so I'm not really expecting it to be an issue).
Changes the format of yaml output to provide more index information,
just as reminders when seeing something like `node+0`. Note this would
create more churn in deltas if we were reliant on the output yaml in
tests, but we aren't so it should be okay.
This standardizes on having ValueStore and related structures provide
printing, removing the handlers in file.cpp.
The `[]` is provided for empty sequences versus if there was simply
nothing, in which case it would be a sequence when non-empty, and a null
value when empty. Consistently (and explicitly) providing sequences
feels easier to understand.
The changes to the output yaml are overall more terse. My hope is that
this is an improvement for most readers.
Also fixes printing of APInt, defaulting to unsigned for consistency
with Carbon's use.
Finishing what #3316 started, add more bespoke ValueStore-like
structures to File. With this, the things which previously had somewhat
boilerplate Add/Get functions are now all on side classes, giving a
uniform style of API for calling.
Note, I was on the fence about making things public on ValueStore. If
it's preferred that I make some things there protected I certainly can,
there's just a trade-off that may mean more distinct child/wrapper
types.
Building on #3313, start using ValueStore on File. Functions and classes
are straightforward. Types here I present as a borderline case where
maybe we want a more bespoke API, but maybe this is okay? Most other
things probably need a slightly different API, which although I might do
that for a consistent interface, felt more out-of-scope for this change.
This replaces the printing that was removed from SemIR's raw dump. It's
separate because (for example) lexing generates shared values, and so
reviewing them is not specific to any particular phase.
Building on #3311, change SemIR to use the SharedValueStore. Since this
removes hermeticity, raw output no longer prints ints, reals, and
strings. TokenizedBuffer accessors are modified to return IDs because
values are often passed through in semantics without needing to read
them.
I would've put SharedValueStores on Context, except for the
GetArrayBoundValue convenience method. I felt awkward removing that, so
it's on File, at least for now. That's then used by the formatter and
Lower too. The flipside of this is that TokenizedBuffer has a
SharedValueStores only for printing, so maybe that's similar enough to
what File is doing.
This doesn't start shifting other SemIR members to ValueStore, but that
seems like a next step.
This updates lexing to use the data. I'll do checking separately, just
to split changes.
Note the ValueStore structure is also set up such that SemIR::File can
use it for other fields.
---------
Co-authored-by: Richard Smith <richard@metafoo.co.uk>
Richard SmithandDana Jansens
Dana Jansens
Chandler Carruth
David Blaikie
Jon Ross-Perkinsandjosh11b
Pablo Paglilla
josh11b